【作者】 刘楠；

【导师】 舒宁；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2005， 硕士

【摘要】 随着遥感技术的进步,高光谱遥感出色的光谱探测能力使得电磁波波谱响应特性更全面准确地反映出地物目标的性质,为遥感影像信息解译提供了良好的平台。遥感数据源在信息提取和识别的过程中主要表现为三大特征:空间特征、光谱特征和时间特征。从信息论的角度分析,空间特征代表了物质信息,它表征物体成分、结构、形状、大小的空间几何特征;光谱特征代表了场信息(能量信息),它表征了物体能量流的成分、结构及特征状态的信息;时间特征代表了动态变化信息,是前两种特征在时间轴上的延续,表征了地物特征的时间分布及与气候环境因素相关的信息。场信息在物质信息的支持下,上升到规则系统和理论化后,构成遥感信息。在成像光谱技术迅猛发展的今天,由于光谱分辨率的不断提高,使得一些传统技术方法的应用遇到困难。成像光谱图像数据特征提取技术的研究,直接关系到针对高维图像数据进行的信息提取和地物识别过程,因此成为这一领域研究中的重要课题,越来越受到遥感工作者的关注。 图像的边缘往往是指图像局部区域亮度变化最显著的部分,该区域的灰度剖面一般可以看作是一个阶跃,即从一个灰度值在很小的缓冲区域内急剧变化到另一个灰度相差较大的灰度值。边缘提取在图像分割、纹理特征提取、形状特征提取、图像识别、计算机视觉等方面的重要性十分突出。虽然传统的微分模板算子和人工神经网络、小波分析、专家系统、数学形态学、遗传算法这些较前沿的理论与方法都在不断地扩大着边缘提取的理论体系,但几乎都不能直接应用到高维影像的处理中来。对高光谱影像而言,此时研究的对象已由象元灰度值转变为灰度矢量,研究域由一维的灰度空间延伸到多维的光谱空间。传统意义上的边缘具有灰度级阶跃变化或屋顶状变化的特性,如今在光谱空间中是否体现出新的特征,以及如何有效地提取这些特征作为判别边缘的依据都是本文研究的重点。 数据场方法是模拟物理学中场论的思想,认为两个数据样本之间存在某种相互作用,通过模拟这种相互作用来发现数据之间内在的联系。本论文在定义了有关数据场的一些基本概念和术语的基础之上,将数据对象的数据辐射思想移植到光谱空间中的光谱矢量特征点上,认为光谱空间中的每个光谱矢量点都相当于一个点电荷或质点,自身带有数据能量,这种能量不仅仅局限于自身点位上,而是向整个光谱空间进行数据辐射,把自身的能量扩散到光谱空间的其他地方从而在其周围形成一个数据场。位于光谱空间中的任何光谱矢量点在向周围辐射自身数据能量的同时,也受到光谱空间中其他光谱矢量点的辐射影响,矢量点与矢量点之间并不是孤立的,而是存在着某种相互制约、相互影响的“作用力”。一个光谱矢量点的特性不仅取决于自身在光谱空间中的位置,在某种程度上更重要的是取决于周围光谱矢量点对该矢量点的辐射强度。周围光谱矢量点的空间分布模式不同(密集或稀疏),该矢量点受到的叠加辐射强度就不同,客观上体现为该矢量点所接受到的辐射能量就不同,不同的能量就体现出不同的特性。这种用数据场中能量的不同来表达一个光谱矢量点的特征就是数据场的构建目的所在。数据场思想把局部区域的光谱矢量点看作一个相互影响相互关联的整体,通过衡量中心象元矢量点在由其邻域光谱矢量点所形成的数据场中反映出的能量特性来分析该中心象元点在现实世界中的实际特性,这也是本文研究的理论基础。最后通过MAIS 30波段影像和TM 7波段影像从不同角度验证了此方法的合理性与可行性。更多还原

【Abstract】 Along with the advancement of Remote Sensing technology, the excellent ability of detecting hyper-spectral remote sensing spectrum makes the spectral response of electromagnetic wave represent the attributes of earth’s surface objects more completely and truly and provide better platform for the interpretation of remote sensed images. During the information detection and discrimination, the remote sensed data always represents three important characters: space character, spectral character and time character. From the analysis of information theory, space character represents material information which exhibits the space geometric features such as the object’s composition, structure, shape and size. And spectral character represents field information which shows the information including objects’ energy and characteristic status. Time information represents the dynamic changing information which continues the space and spectral information and also includes surface features’ time distribution and environmental factors about the climate. Based on material information, field information constructs remote sensed information after forming the regular system and method. Today, because of the constant increase of spectral resolution, it is difficult to the application of traditional image processing methods. The research of the feature extraction of hyper-spectral remote sensing data will be a very important way because it is about the process of information extraction and object identification.Image edge often is the most remarkable part of gray variance of image local area whose gray section may be seen a skip where the gray changes in a small area from low gray level to high gray level. Edge detection is very significant in the domain of image fragmentation, texture detection, computer view and so on. Although some traditional methods, such as differential template operators, artificial nervous network, wavelet analysis, expert system, mathematic morphology, genetic algorithm, are expanding the method system of edge detection, they almost have no ability to apply in the hyper-spectral data directly. For hyper-spectral images, the research contents have been changed from gray value to gray vector and from one-dimensional space to hyper-dimensional space. It is valuable to discover whether in the spectral space there are some new properties of edge points which can be regarded as gray non-continuous change in former viewpoint, and the way how to detect such properties.Data field is a kind of method of simulating field theory of physics. The main idea is that two data samples have some interaction which could be applied to represent the inner relationship between two data. Based on several fundamental concepts and terms, the idea of data radiation would be transplanted in the spectral space. Data radiation views any spectral vector point of the spectral space as an electric charge which owns some energy forming a data filed through radiating to the whole spectral space in order to affect other surrounding spectral vector points. The spectral vector point not only radiates its own energy to other points, but also receives the energy radiated from other vector points around it. There exits some invisible power between vector points. These points arenot lonely. The property of a vector point would not only been determined by the position in the spectral space, to some extent more importantly, by the radiation intensity of its surrounding vector points. The difference of distribution mode of surrounding spectral vector points makes the superposed radiation and energy different which represents the property of vector point different. It is the final goal to use the different energy of point in the data field presents the different property of the point. Data field theory views the local spectral vector points as a group in which every element interacts each other. The base of this paper is that a point’s true property can be determined through calculating the field energy of point in a data field generated by this point’s surrounding local vector points. Finally, we verify the reasonability and feasibility of the data field methods through MAIS 30 bands remote sensed images and TM 7 bands remote sensed images.更多还原